<Title> Yiqun Hu, SP Group Agenda Condition monitoring - - PowerPoint PPT Presentation

<Title>

Autoencoder Forest for Anomaly Detection from IoT Time Series

Yiqun Hu, SP Group

Data Council Singapre 2019

Agenda

• Condition monitoring & anomaly detection
• Autoencoder for anomaly detection
• Autoencoder Forest
• End-to-end workflow
• Experiment results

Conditional monitoring & Anomaly Detection

Condition monitoring

• Manual monitoring

– Huge human effort – Boring task with low quality

• Rule-based method

– Cannot differentiate different environment – Cannot adapt to different condition of the equipment

• Data-driven method

– Model the common behavior of the equipment

Time-series anomaly detection

Autoencoder for Anomaly Detection

Autoencoder

• What is autoencoder

– A encoder-decoder type of neural network architecture that is used for self-learning from unlabeled data

• The idea of autoencoder

– Learn how to compress data into a concise representation to allow for the reconstruction with minimum error

• Different variants of autoencoder

– Variational Autoencoder – LSTM Autoencoder – Etc.

Autoencoder Neural Network

Autoencoder for anomaly detection

Online Detection Anomaly score Offline Training Reconstruction errors

Autoencoder Forest

A key challenge of autoencoder

Single Autoencoder

The idea of autoencoder forest

x x xx x x x x x

• o o o o
• o
• ++

+ + + + +

Clustering subsequence is meaningless

[1]. Eamonn Keogh, Jessica Lin, Clustering of Time Series Subsequences is Meaningless: Implications for Previous and Future Research

Autoencoder forest based on time

0:00 1:00 1:30 22:00 23:30

Training autoencoder forest

Input Layer Encoder layer 1

(window_size, 1) (window_size/2, 1)

Encoder layer 2

(window_size/4, 1)

Decoder layer 1

(window_size/2, 1)

Decoder Layer 2

(window_size, 1)

• Structure is fixed for every
• autoencoder. (try to make it

as generic as possible)

• Each autoencoder within

forest is independent. So the training is naturally parallelizable

• Using early stopping

mechanism, the training of individual autoencoder can be stopped at similar accuracy.

Autoencoder Forest

Single Autoencoder Autoencoder Forest

End-to-end Workflow

Automatic end-to-end workflow

Time series analysis Train Data Preprocessing Train Window Extraction Autoencoder Forest Training Test Data Preprocessing Test Window Extraction Anomaly scoring

Training Anomaly detection

Periodic pattern analysis

• Automatic determine the

repeating period in time series

– Calculate autocorrelations of different lags – Find the strong local maximum

• f autocorrelation

– Calculate the interval of any two local maximum – Find the mode of intervala

Missing data handling

3:05 3:10 3:15 3:20 … … 16:15 16:21 16:24 16:30

… … Misalignment Missing

3:05 3:10 3:15 3:20 … … 16:15 (16:20 – 16:40) 16:45

… … ? ? ?

• No need to impute
• If missing gap is small,

impute with neighbouring points;

• If missing gap is large,

impute with the same time of other periods;

Anomaly scoring

Extract the sequence window end at time t

. . . . . .

Median profile

Corresponding autoencoder reconstruct the sequence window at time t Compute reconstruction error as anomaly score Learned autoencoder forest

Experiment Results

Cooling tower – return water temperature

Chiller – chilled water return temperature

Smart meter – half hour consumption

2018-12-03 22:00:00 Normal data 2018-09-27 14:30:00 2018-10-06 22:30:00 2018-09-07 15:30:00 Top 3 Detected Anomaly

A common platform for time series data, with built-in AI capabilities

powering the nation